def output_predictions(dataset: Dataset, y_preds: np.ndarray,
csv_out: str) -> None:
"""Writes predictions to file.
Writes predictions made on `dataset` to a specified file on
disk. `dataset.ids` are used to format predictions. The produce CSV file will have format as follows
| ID | Task1Name | Task2Name |
| ----------- | ------------ | ------------ |
| identifer1 | prediction11 | prediction12 |
| identifer2 | prediction21 | prediction22 |
Parameters
----------
dataset: dc.data.Dataset
Dataset on which predictions have been made.
y_preds: np.ndarray
Predictions to output
csv_out: str
Name of file to write predictions to.
"""
data_ids = dataset.ids
n_tasks = len(dataset.get_task_names())
y_preds = np.reshape(y_preds, (len(y_preds), n_tasks))
assert len(y_preds) == len(data_ids)
with open(csv_out, "w") as csvfile:
csvwriter = csv.writer(csvfile)
csvwriter.writerow(["ID"] + dataset.get_task_names())
for mol_id, y_pred in zip(data_ids, y_preds):
csvwriter.writerow([mol_id] + list(y_pred))
def predict(self,
dataset: Dataset,
transformers: List[Transformer] = []) -> np.ndarray:
"""
Uses self to make predictions on provided Dataset object.
Parameters
----------
dataset: Dataset
Dataset to make prediction on
transformers: List[Transformer]
Transformers that the input data has been transformed by. The output
is passed through these transformers to undo the transformations.
Returns
-------
np.ndarray
A numpy array of predictions the model produces.
"""
y_preds = []
for (X_batch, _, _,
ids_batch) in dataset.iterbatches(deterministic=True):
n_samples = len(X_batch)
y_pred_batch = self.predict_on_batch(X_batch)
# Discard any padded predictions
y_pred_batch = y_pred_batch[:n_samples]
y_pred_batch = undo_transforms(y_pred_batch, transformers)
y_preds.append(y_pred_batch)
y_pred = np.concatenate(y_preds)
return y_pred
def fit(self, dataset: Dataset, nb_epoch: int = 10) -> float:
"""
Fits a model on data in a Dataset object.
Parameters
----------
dataset: Dataset
the Dataset to train on
nb_epoch: int
the number of epochs to train for
Returns
-------
float
The average loss over the most recent checkpoint interval.
"""
for epoch in range(nb_epoch):
logger.info("Starting epoch %s" % str(epoch + 1))
losses = []
for (X_batch, y_batch, w_batch,
ids_batch) in dataset.iterbatches():
losses.append(self.fit_on_batch(X_batch, y_batch, w_batch))
logger.info("Avg loss for epoch %d: %f" %
(epoch + 1, np.array(losses).mean()))
return np.array(losses).mean()
def predict(self,
dataset: Dataset,
transformers: List[Transformer] = []) -> OneOrMany[np.ndarray]:
"""
Uses self to make predictions on provided Dataset object.
Parameters
----------
dataset: dc.data.Dataset
Dataset to make prediction on
transformers: list of dc.trans.Transformers
Transformers that the input data has been transformed by. The output
is passed through these transformers to undo the transformations.
Returns
-------
a NumPy array of the model produces a single output, or a list of arrays
if it produces multiple outputs
"""
y_preds = []
n_tasks = self.get_num_tasks()
ind = 0
for (X_batch, _, _,
ids_batch) in dataset.iterbatches(deterministic=True):
n_samples = len(X_batch)
y_pred_batch = self.predict_on_batch(X_batch)
# Discard any padded predictions
y_pred_batch = y_pred_batch[:n_samples]
y_pred_batch = undo_transforms(y_pred_batch, transformers)
y_preds.append(y_pred_batch)
y_pred = np.concatenate(y_preds)
return y_pred
def fit(self, dataset: Dataset, nb_epoch: int = 10) -> float:
"""
Fits a model on data in a Dataset object.
Parameters
----------
dataset: Dataset
the Dataset to train on
nb_epoch: int
the number of epochs to train for
Returns
-------
the average loss over the most recent epoch
"""
# TODO(rbharath/enf): We need a structured way to deal with potential GPU
# memory overflows.
for epoch in range(nb_epoch):
logger.info("Starting epoch %s" % str(epoch + 1))
losses = []
for (X_batch, y_batch, w_batch,
ids_batch) in dataset.iterbatches():
losses.append(self.fit_on_batch(X_batch, y_batch, w_batch))
logger.info("Avg loss for epoch %d: %f" %
(epoch + 1, np.array(losses).mean()))
return np.array(losses).mean()
def __init__(self, dataset: Dataset):
super(DuplicateBalancingTransformer, self).__init__(transform_X=True,
transform_y=True,
transform_w=True,
transform_ids=True,
dataset=dataset)
if len(dataset.get_task_names()) > 1:
raise ValueError(
"This transformation is only defined for singletask datsets.")
# Get the labels/weights
y = dataset.y
w = dataset.w
# Normalize shapes
if len(y.shape) == 1:
y = np.reshape(y, (len(y), 1))
if len(w.shape) == 1:
w = np.reshape(w, (len(w), 1))
if len(y.shape) != 2:
raise ValueError("y must be of shape (N,) or (N, n_tasks)")
if len(w.shape) != 2:
raise ValueError("w must be of shape (N,) or (N, n_tasks)")
self.classes = sorted(np.unique(y))
# Remove labels with zero weights
y = y[w != 0]
N = len(y)
class_weights = []
# Note that we may have 0 elements of a given class since we remove those
# labels with zero weight.
for c in self.classes:
# this works because y is 1D
c_weight = np.sum(w[y == c])
class_weights.append(c_weight)
weight_largest = max(class_weights)
# This is the right ratio since int(N/num_c) * num_c \approx N
# for all classes
duplication_ratio = [
int(weight_largest / float(c_weight)) if c_weight > 0 else 0
for c_weight in class_weights
]
self.duplication_ratio = duplication_ratio
def default_generator(
self,
dataset: Dataset,
epochs: int = 1,
mode: str = 'fit',
deterministic: bool = True,
pad_batches: bool = True) -> Iterable[Tuple[List, List, List]]:
"""Create a generator that iterates batches for a dataset.
Subclasses may override this method to customize how model inputs are
generated from the data.
Parameters
----------
dataset: Dataset
the data to iterate
epochs: int
the number of times to iterate over the full dataset
mode: str
allowed values are 'fit' (called during training), 'predict' (called
during prediction), and 'uncertainty' (called during uncertainty
prediction)
deterministic: bool
whether to iterate over the dataset in order, or randomly shuffle the
data for each epoch
pad_batches: bool
whether to pad each batch up to this model's preferred batch size
Returns
-------
a generator that iterates batches, each represented as a tuple of lists:
([inputs], [outputs], [weights])
"""
for epoch in range(epochs):
for (X_b, y_b, w_b, ids_b) in dataset.iterbatches(
batch_size=self.batch_size,
deterministic=deterministic,
pad_batches=pad_batches):
yield ([X_b], [y_b], [w_b])
def default_generator(
self,
dataset: Dataset,
epochs: int = 1,
mode: str = 'fit',
deterministic: bool = True,
pad_batches: bool = True) -> Iterable[Tuple[List, List, List]]:
"""Convert a dataset into the tensors needed for learning.
Parameters
----------
dataset: `dc.data.Dataset`
Dataset to convert
epochs: int, optional (Default 1)
Number of times to walk over `dataset`
mode: str, optional (Default 'fit')
Ignored in this implementation.
deterministic: bool, optional (Default True)
Whether the dataset should be walked in a deterministic fashion
pad_batches: bool, optional (Default True)
If true, each returned batch will have size `self.batch_size`.
Returns
-------
Iterator which walks over the batches
"""
for epoch in range(epochs):
for (X_b, y_b, w_b, ids_b) in dataset.iterbatches(
batch_size=self.batch_size,
deterministic=deterministic,
pad_batches=pad_batches):
if y_b is not None:
if self.mode == 'classification':
y_b = to_one_hot(y_b.flatten(), self.n_classes).reshape(
-1, self.n_tasks, self.n_classes)
inputs = self.compute_features_on_batch(X_b)
yield (inputs, [y_b], [w_b])
def default_generator(
self,
dataset: Dataset,
epochs: int = 1,
mode: str = 'fit',
deterministic: bool = True,
pad_batches: bool = True) -> Iterable[Tuple[List, List, List]]:
"""Convert a dataset into the tensors needed for learning.
Parameters
----------
dataset: `dc.data.Dataset`
Dataset to convert
epochs: int, optional (Default 1)
Number of times to walk over `dataset`
mode: str, optional (Default 'fit')
Ignored in this implementation.
deterministic: bool, optional (Default True)
Whether the dataset should be walked in a deterministic fashion
pad_batches: bool, optional (Default True)
If true, each returned batch will have size `self.batch_size`.
Returns
-------
Iterator which walks over the batches
"""
for epoch in range(epochs):
for (X_b, y_b, w_b, ids_b) in dataset.iterbatches(
batch_size=self.batch_size,
deterministic=deterministic,
pad_batches=pad_batches):
if y_b is not None:
if self.mode == 'classification':
y_b = to_one_hot(y_b.flatten(), self.n_classes).reshape(
-1, self.n_tasks, self.n_classes)
atom_feat = []
pair_feat = []
atom_split = []
atom_to_pair = []
pair_split = []
start = 0
for im, mol in enumerate(X_b):
n_atoms = mol.get_num_atoms()
# number of atoms in each molecule
atom_split.extend([im] * n_atoms)
# index of pair features
C0, C1 = np.meshgrid(np.arange(n_atoms), np.arange(n_atoms))
atom_to_pair.append(
np.transpose(
np.array([C1.flatten() + start,
C0.flatten() + start])))
# number of pairs for each atom
pair_split.extend(C1.flatten() + start)
start = start + n_atoms
# atom features
atom_feat.append(mol.get_atom_features())
# pair features
pair_feat.append(
np.reshape(mol.get_pair_features(),
(n_atoms * n_atoms, self.n_pair_feat[0])))
inputs = [
np.concatenate(atom_feat, axis=0),
np.concatenate(pair_feat, axis=0),
np.array(pair_split),
np.array(atom_split),
np.concatenate(atom_to_pair, axis=0)
]
yield (inputs, [y_b], [w_b])
def load_bace(mode="regression", transform=True, split="20-80"):
"""Load BACE-1 dataset as regression/classification problem."""
reload = True
verbosity = "high"
regen = False
assert split in ["20-80", "80-20"]
current_dir = os.path.dirname(os.path.realpath(__file__))
if split == "20-80":
dataset_file = os.path.join(current_dir,
"../../datasets/desc_canvas_aug30.csv")
elif split == "80-20":
dataset_file = os.path.join(current_dir,
"../../datasets/rev8020split_desc.csv")
dataset = load_from_disk(dataset_file)
num_display = 10
pretty_columns = ("[" + ",".join(
["'%s'" % column
for column in dataset.columns.values[:num_display]]) + ",...]")
crystal_dataset_file = os.path.join(
current_dir, "../../datasets/crystal_desc_canvas_aug30.csv")
crystal_dataset = load_from_disk(crystal_dataset_file)
print("Columns of dataset: %s" % pretty_columns)
print("Number of examples in dataset: %s" % str(dataset.shape[0]))
print("Number of examples in crystal dataset: %s" %
str(crystal_dataset.shape[0]))
#Make directories to store the raw and featurized datasets.
base_dir = tempfile.mkdtemp()
data_dir = os.path.join(base_dir, "dataset")
train_dir = os.path.join(base_dir, "train_dataset")
valid_dir = os.path.join(base_dir, "valid_dataset")
test_dir = os.path.join(base_dir, "test_dataset")
model_dir = os.path.join(base_dir, "model")
crystal_dir = os.path.join(base_dir, "crystal")
if mode == "regression":
bace_tasks = ["pIC50"]
elif mode == "classification":
bace_tasks = ["Class"]
else:
raise ValueError("Unknown mode %s" % mode)
featurizer = UserDefinedFeaturizer(user_specified_features)
loader = DataLoader(tasks=bace_tasks,
smiles_field="mol",
id_field="CID",
featurizer=featurizer)
if not reload or not os.path.exists(data_dir):
dataset = loader.featurize(dataset_file, data_dir)
regen = True
else:
dataset = Dataset(data_dir, reload=True)
if not reload or not os.path.exists(crystal_dir):
crystal_dataset = loader.featurize(crystal_dataset_file, crystal_dir)
else:
crystal_dataset = Dataset(crystal_dir, reload=True)
if (not reload or not os.path.exists(train_dir)
or not os.path.exists(valid_dir) or not os.path.exists(test_dir)):
regen = True
splitter = SpecifiedSplitter(dataset_file,
"Model",
verbosity=verbosity)
train_dataset, valid_dataset, test_dataset = splitter.train_valid_test_split(
dataset, train_dir, valid_dir, test_dir)
else:
train_dataset = Dataset(train_dir, reload=True)
valid_dataset = Dataset(valid_dir, reload=True)
test_dataset = Dataset(test_dir, reload=True)
#NOTE THE RENAMING:
if split == "20-80":
valid_dataset, test_dataset = test_dataset, valid_dataset
print("Number of compounds in train set")
print(len(train_dataset))
print("Number of compounds in validation set")
print(len(valid_dataset))
print("Number of compounds in test set")
print(len(test_dataset))
print("Number of compounds in crystal set")
print(len(crystal_dataset))
if transform and regen:
input_transformers = [
NormalizationTransformer(transform_X=True, dataset=train_dataset),
ClippingTransformer(transform_X=True, dataset=train_dataset)
]
output_transformers = []
if mode == "regression":
output_transformers = [
NormalizationTransformer(transform_y=True,
dataset=train_dataset)
]
else:
output_transformers = []
else:
input_transformers, output_transformers = [], []
transformers = input_transformers + output_transformers
for dataset in [
train_dataset, valid_dataset, test_dataset, crystal_dataset
]:
for transformer in transformers:
dataset = transformer.transform(dataset)
return (bace_tasks, train_dataset, valid_dataset, test_dataset,
crystal_dataset, output_transformers)
